Early access · building in the open

The control plane for AI engineering.

Your AI coding agents — Claude Code, Codex, Cursor — do the work in sandboxes. FlightDeck owns the mission, the policy, the verification, and the record. One governed loop over every vendor.

AIREC
Powered by AIREC.

Every FlightDeck mission rides AIREC's hash-chained evidence spine — the same record your governance and security teams already trust.

Explore the platform →
The problem

AI agents are great at writing code. Terrible at being trusted with it.

Single-vendor agents make you choose one and hope. FlightDeck governs them all.

Every agent owns its own loop

It decides, edits, and "finishes" with no oversight layer above it. You find out what it did afterward.

No isolation

Most agents work directly in your real tree. A bad run touches the code that matters before you can stop it.

No record across vendors

Codex today, Cursor tomorrow — and nothing carries the mission, the decisions, or the proof between them.

The governed loop

Dispatch. Watch. Verify. Accept.

One loop wraps every agent. The work happens in a sandbox; nothing reaches your real tree until you say so — and the whole run is a replayable record.

01 · DISPATCH

Into a sandbox

FlightDeck spins up a sandboxed git worktree for the agent.

🔒 isolated from main
02 · WATCH

Live, in the console

The agent runs headless inside FlightDeck; its output streams in the Run Console. No terminal, no copy-paste.

03 · VERIFY

Your checks, in-sandbox

Your committed tests run inside the agent's sandbox — green before you even look.

✓ verified in isolation
04 · ACCEPT

Only on your click

Nothing merges until you accept — and the whole mission is a hash-chained, replayable record.

⤷ recorded as canon
"Codex worked in a sandbox, FlightDeck verified its tests, nothing touched main until I clicked accept — and here's the record."
Multi-vendor by design

Bring your own agents.

FlightDeck speaks their protocol (MCP) and governs them through one control plane. Add a vendor; the governance doesn't change.

▲ FlightDeck control plane — one governed loop
Claude Codevia MCP
Codexvia MCP
Cursorvia MCP
3 vendors — Claude Code · Codex · Cursor
10 MCP tools the agents call
Devin, Copilot Workspace, Cursor's agents each own their own loop — FlightDeck owns the loop above all of them
The record

Every mission is evidence.

Not "done" — a structured, reviewable record of exactly what each agent did, linked end to end.

Hash-chained mission log

Every dispatch, signal, and decision linked in a tamper-evident chain.

Structured agent returns

Files changed, commands run, tests, risk flags — not just a thumbs-up.

Evidence packs

Export a signed bundle on AIREC's spine and verify it offline.

Authority Layer (v0.1)

When two agents disagree, you resolve it — and your decision is recorded as canon.

Where it's going

Shipping in the open, with our design partners.

Here's what's on the roadmap — not yet shipped. Early access means you help shape what lands and in what order.

Automatic agent routing by task type
Mission decomposition
Idea-to-roadmap studio
Merge gates
Cross-agent conflict arbitration
Per-artifact Ed25519 signing
Today — agent routing is manual defaults; conflicts are human-arbitrated
Honest about what's shipped vs. ahead — that's the point of early access